Instead of total monthly sales (S) as your dependent variable, you could model average monthly sales per customer (S/C).  You could also examine the independent variables added to the regression on S of C that change the sign of the regression coefficient of C.  What combination of these independent variables change this sign and why?  Perhaps such an examination would identify subgroups based on levels of these independent variables or interactions between C and these independent variables that might be worth further study and modelling. ... View more

Read the following reference that explains differences between design-based analyses like those in the SAS survey procedures and model-based analyses like those using multilevel mixed models (for example, PROC MIXED or PROC GLIMMIX):    Graubard BI, Korn EL.  Modeling the sampling design in the analysis of health surveys.  Statistical        Methods in Medical Research 1996;5:263-281. Sometimes the variance of parameters (for example, means) from design-based analyses are less than that from model-based analyses, but other times the reverse is true.  For example, as the number of primary sampling units increase, the variances estimated during design-based analyses are approximately unbiased but may be biased if this number is small.   If the values or the variability of a parameter are associated with the number of observations in a primary sampling unit, then the variances of the parameter estimated during model-based analyses may be biased. The authors generally prefer design-based analyses to model-based analyses because the former requires fewer assumptions than the latter. The characteristics of the survey design in design-based analyses (the strata and the primary sampling units) are considered "nuisances" and of little interest for the analyst.  owever, if these characteristics of the survey design may be associated with the parameters of interest (as above), and if these characteristics are available to the analyst (which is uncommon), then modeling that accounts for these characteristics may be worthwhile (for example, in reducing the bias of model-based estimates of variance). Although this article does not specifically compare design-based analyses with analyses based on generalized estimating equations, the same considerations probably apply. ... View more

I don't have an answer to your problem.  Your problem requires that you specify what you mean by persons being similar to one another. If you have 34 choices, and if you select 5 of them to rank from 1 to 5, the total number of such ranked choices is the number of permutations of 34 items taken 5 at a time = 34!/(34-5)! = 34*33*32*31*30 = 33,390,720.  Because these permutations preserve the order of the ranked choices, two persons could select the same set of five items but rank each of these items differently so that they would be similar/identical in their choices but not similar in how they ranked these choices.  Thus, you should specify how you determine how someone is "similar" to someone else and which similarity criterion is most important to you--similarity in the choices or similarity in the ranking of those choices.  For example, persons A and B could have selected the same five choices but have ranked them in an exact opposite order; person C could have selected four of the same five choices as person B and ranked those four choices identically to person B's ranking of these four choices.  Is person A more similar than person C to person B?  Or, vice-versa? ... View more

Check the following reference at Lex Jansen's Internet site:   http://www.lexjansen.com/wuss/2012/162.pdf This reference shows several methods on how to adjust the sampling weights from the observed responses in your survey to conform with the population totals you want to poststratify to.  Poststratification implies an external standard population that provides these population totals, but such a population is not necessary to adjust for survey nonresponse.  Then use these new poststratified weights in the SAS survey analysis procedures. ... View more

I think the discrepancy is due to your use in the STATA SVYSET command of the options, POSTSTRATA and POSTWEIGHT.  These options are used to adjust the respondent sampling weights so that they sum to the population sizes within each poststratum to account for nonresponse and underrepresented groups in the population (cf., the STATA documentation).  These poststratification strata differ from the "design" strata used in your complex sample survey and in the SAS STRATA statement.  In SAS, these poststratification adjustments are usually performed beforehand on the respondent sampling weights so that these poststratified, adjusted respondent sampling weights are used in the SAS WEIGHT statement. Note that your STATA output implies that the DESIGN of your survey has only one stratum and 4,337 primary sampling units.  STATA lists the number of postrata as 17.  However, since your SAS syntax uses the variable, POSTSTRATA, as the argument of its STRATA statement, the DESIGN of your survey in SAS implies 17 strata and 4,337 primary sampling units.  Thus, STATA "sees" only one stratum in your sample design, and SAS "sees" 17 strata in your sample design. To make the STATA output conform with the SAS output, change the STATA SVYSET command to specify 17 strata and 4,337 primary sampling units:     SVYSET, CLEAR      SVYSET _n [pweight=aweight], strata(poststrata)   ... View more

Were the weights integral weights equal to 1.00?  If not, why would you expected a weighted sum equal to an integral number of observations particularly in a DOMAIN analysis where you are studying a subgroup of the entire sample? ... View more

Your original question asked why the results from SAS were not the same as those from STATA.  Since you did not show the results from either program, all I could do was suggest one possible reason why these results may differ.  Although changing the reference group for INTERVIEWMETHOD should not change the overall effect of deprivation after adjustment, if you had changed this reference group, did the results from SAS still differ from those of STATA? ... View more

Obviously, that didn't work.  When I tried your PROC IML program you sent, the program took a very long time and did NOT converge.  I tried reducing the number of iterations and changing the input data, but the problem seems to be with the slow convergence of the Nelder-Mead simplex method on your problem.  Do you have previously published input data for your problem that are known to converge to known solutions using this method?  If so, try to see if you obtain these solutions using these input data.  Otherwise, see if you can recast your problem so that another optimization method might work with it.  What is the purpose of this optimization? ... View more

By default, STATA's method of creating indicator ("dummy") variables for a categorical variable is to create such variables for all levels of the categorical variable and to omit the indicator variable corresponding to the smallest level of the categorical variable.  One of your categorical variables in the SAS code, INTERVIEWMETHOD, has a reference level [=3] that is probably NOT the smallest level for this variable [probably something less than 3].  Therefore, to translate the STATA coding to SAS, specify the smallest level for INTERVIEWMETHOD as its reference level in the CLASS statement. Or, you can study the STATA documentation to change the default behavior for STATA in creating indicator variables to specify instead level 3 of INTERVIEWMETHOD as its reference level. ... View more

Could you specify your two-parameter nonlinear model? According to your output,                                                              X1                          X2        Value of objective function               Starting parameter values:       0.798                      0.150                    3.4654     Optimized parameter values:   0                              0.880851             3.45307 The constraint violation occurs only with the optimized value of X1, which is minimally different from zero but negative (not meeting your lower bound constraint).  However, this violation of the constraint is minimal. The facts that the optimized values of X1 and X2 differ so much from the starting parameter values, that the optimized value for one of these values is on the constraint boundary, and that the value of the objective function changes so minimally imply that many parameter values are consistent with your data, that the objective function "space" is basically flat, and that you probably need more data to distinguish among these values to find a global minimum.  What happens to the values of the objective function when you remove or widen the constraints? ... View more

The default coding for categorical variables in the CLASS statement (for example, Treatment) in PROC LOGISTIC is effect coding.  Preferable is the use of reference coding, which estimates the difference in the effect of each nonreference level to the effect of the reference level.  The difficulty with your example data is that either of your two "control" treatment levels (1 or 2) could be chosen as the reference level.  What I usually do to force the choice of a reference level is to include a prior PROC FORMAT paragraph VALUE statement for the classification variable that specifies this reference level:      proc format;          value treatfm 1="} Treatment 1 (ref)"                              2="Treatment 2"                              3="Treatment 3"      run;      proc logistic data=Mydata;          class Treatment(order=formatted param=reference);          model Y = Treatment / clodds=both;          format Treatment treatfm.;      run; The right brace ["}"] in the formatted value for Treatment 1 in the VALUE statement of PROC FORMAT will sort this value last (ORDER=FORMATTED) and thus specifies this value as the reference level for the categorical variable, Treatment.  Reference coding like this may not cause the anomaly that you describe with a not statistically significant "P-value" occurring with 95% confidence intervals for the odds ratio that exclude the null value of 1.00. Another approach to estimating odds ratios would be to use a CONTRAST statement in PROC LOGISTIC, a statement like the ESTIMATE statement you used in PROC GLIMMIX.   If the mean of a binomial distribution estimates the proportion of Y=1 outcomes across all the observations, the variance would estimate the variability of these proportions across the observations and could be used to estimate a confidence interval around that mean, which resembles what you can do in the continuous case (mean and confidence interval around that mean). The SOLUTION option of the RANDOM statement estimates the random intercept for the observation(s) identified in the SUBJECT option.  If SUBJECT=Animal_ID identifies one or more observations for one animal, this random intercept is the mean estimate of that value across these observations.  This value, along with the values estimated from the fixed effects, can be used to estimate the linear predictor ("best linear unbiased predictor") in the OUTPUT statement of PROC GLIMMIX.  Although this random estimate is "constant" within each SUBJECT=Animal_ID, differences in the fixed effects among the possibly multiple observations associated with each Animal_ID would yield different best linear unbiased predictors for these different observations:      best linear unbiased predictor(Y) = b0[=fixed effect intercept] + b1*(Treatment=2) + b2*(Treatment=3) + z0i[=random effect intercept for Animal_ID]          where b0, b1, and b2=fixed effects, and z0i=random intercept for SUBJECT=Animal_IDi. ... View more